1 /* Copyright (c) 2007-2017. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
7 * Modeling the proportional fairness using the Lagrangian Optimization Approach. For a detailed description see:
8 * "ssh://username@scm.gforge.inria.fr/svn/memo/people/pvelho/lagrange/ppf.ps".
10 #include "src/kernel/lmm/maxmin.hpp"
12 #include "xbt/sysdep.h"
20 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(surf_lagrange, surf, "Logging specific to SURF (lagrange)");
21 XBT_LOG_NEW_SUBCATEGORY(surf_lagrange_dichotomy, surf_lagrange, "Logging specific to SURF (lagrange dichotomy)");
23 #define SHOW_EXPR(expr) XBT_CDEBUG(surf_lagrange, #expr " = %g", expr);
24 #define VEGAS_SCALING 1000.0
25 #define RENO_SCALING 1.0
26 #define RENO2_SCALING 1.0
32 double (*func_f_def)(const s_lmm_variable_t&, double);
33 double (*func_fp_def)(const s_lmm_variable_t&, double);
34 double (*func_fpi_def)(const s_lmm_variable_t&, double);
37 * Local prototypes to implement the Lagrangian optimization with optimal step, also called dichotomy.
39 // solves the proportional fairness using a Lagrangian optimization with dichotomy step
40 void lagrange_solve(lmm_system_t sys);
41 // computes the value of the dichotomy using a initial values, init, with a specific variable or constraint
42 static double dichotomy(double init, double diff(double, void*), void* var_cnst, double min_error);
43 // computes the value of the differential of constraint param_cnst applied to lambda
44 static double partial_diff_lambda(double lambda, void* param_cnst);
46 static int __check_feasible(xbt_swag_t cnst_list, xbt_swag_t var_list, int warn)
51 xbt_swag_t elem_list = nullptr;
52 lmm_element_t elem = nullptr;
53 lmm_constraint_t cnst = nullptr;
54 lmm_variable_t var = nullptr;
56 xbt_swag_foreach(_cnst, cnst_list)
58 cnst = static_cast<lmm_constraint_t>(_cnst);
60 elem_list = &(cnst->enabled_element_set);
61 xbt_swag_foreach(_elem, elem_list)
63 elem = static_cast<lmm_element_t>(_elem);
65 xbt_assert(var->sharing_weight > 0);
69 if (double_positive(tmp - cnst->bound, sg_maxmin_precision)) {
71 XBT_WARN("The link (%p) is over-used. Expected less than %f and got %f", cnst, cnst->bound, tmp);
74 XBT_DEBUG("Checking feasability for constraint (%p): sat = %f, lambda = %f ", cnst, tmp - cnst->bound,
78 xbt_swag_foreach(_var, var_list)
80 var = static_cast<lmm_variable_t>(_var);
81 if (not var->sharing_weight)
85 XBT_DEBUG("Checking feasability for variable (%p): sat = %f mu = %f", var, var->value - var->bound, var->mu);
87 if (double_positive(var->value - var->bound, sg_maxmin_precision)) {
89 XBT_WARN("The variable (%p) is too large. Expected less than %f and got %f", var, var->bound, var->value);
96 static double new_value(lmm_variable_t var)
100 for (s_lmm_element_t const& elem : var->cnsts) {
101 tmp += elem.constraint->lambda;
105 XBT_DEBUG("\t Working on var (%p). cost = %e; Weight = %e", var, tmp, var->sharing_weight);
106 // uses the partial differential inverse function
107 return var->func_fpi(*var, tmp);
110 static double new_mu(lmm_variable_t var)
113 double sigma_i = 0.0;
115 for (s_lmm_element_t const& elem : var->cnsts) {
116 sigma_i += elem.constraint->lambda;
118 mu_i = var->func_fp(*var, var->bound) - sigma_i;
124 static double dual_objective(xbt_swag_t var_list, xbt_swag_t cnst_list)
128 lmm_constraint_t cnst = nullptr;
129 lmm_variable_t var = nullptr;
133 xbt_swag_foreach(_var, var_list)
135 var = static_cast<lmm_variable_t>(_var);
136 double sigma_i = 0.0;
138 if (not var->sharing_weight)
141 for (s_lmm_element_t const& elem : var->cnsts)
142 sigma_i += elem.constraint->lambda;
147 XBT_DEBUG("var %p : sigma_i = %1.20f", var, sigma_i);
149 obj += var->func_f(*var, var->func_fpi(*var, sigma_i)) - sigma_i * var->func_fpi(*var, sigma_i);
152 obj += var->mu * var->bound;
155 xbt_swag_foreach(_cnst, cnst_list)
157 cnst = static_cast<lmm_constraint_t>(_cnst);
158 obj += cnst->lambda * cnst->bound;
164 void lagrange_solve(lmm_system_t sys)
166 /* Lagrange Variables. */
167 int max_iterations = 100;
168 double epsilon_min_error = 0.00001; /* this is the precision on the objective function so it's none of the
169 configurable values and this value is the legacy one */
170 double dichotomy_min_error = 1e-14;
171 double overall_modification = 1;
173 XBT_DEBUG("Iterative method configuration snapshot =====>");
174 XBT_DEBUG("#### Maximum number of iterations : %d", max_iterations);
175 XBT_DEBUG("#### Minimum error tolerated : %e", epsilon_min_error);
176 XBT_DEBUG("#### Minimum error tolerated (dichotomy) : %e", dichotomy_min_error);
178 if (XBT_LOG_ISENABLED(surf_lagrange, xbt_log_priority_debug)) {
182 if (not sys->modified)
185 /* Initialize lambda. */
186 xbt_swag_t cnst_list = &(sys->active_constraint_set);
188 xbt_swag_foreach(_cnst, cnst_list)
190 lmm_constraint_t cnst = (lmm_constraint_t)_cnst;
192 cnst->new_lambda = 2.0;
193 XBT_DEBUG("#### cnst(%p)->lambda : %e", cnst, cnst->lambda);
197 * Initialize the var list variable with only the active variables.
198 * Associate an index in the swag variables. Initialize mu.
200 xbt_swag_t var_list = &(sys->variable_set);
202 xbt_swag_foreach(_var, var_list)
204 lmm_variable_t var = static_cast<lmm_variable_t>(_var);
205 if (not var->sharing_weight)
208 if (var->bound < 0.0) {
209 XBT_DEBUG("#### NOTE var(%p) is a boundless variable", var);
215 var->value = new_value(var);
216 XBT_DEBUG("#### var(%p) ->weight : %e", var, var->sharing_weight);
217 XBT_DEBUG("#### var(%p) ->mu : %e", var, var->mu);
218 XBT_DEBUG("#### var(%p) ->weight: %e", var, var->sharing_weight);
219 XBT_DEBUG("#### var(%p) ->bound: %e", var, var->bound);
220 auto weighted = std::find_if(begin(var->cnsts), end(var->cnsts),
221 [](s_lmm_element_t const& x) { return x.consumption_weight != 0.0; });
222 if (weighted == end(var->cnsts))
227 /* Compute dual objective. */
228 double obj = dual_objective(var_list, cnst_list);
230 /* While doesn't reach a minimum error or a number maximum of iterations. */
232 while (overall_modification > epsilon_min_error && iteration < max_iterations) {
234 XBT_DEBUG("************** ITERATION %d **************", iteration);
235 XBT_DEBUG("-------------- Gradient Descent ----------");
237 /* Improve the value of mu_i */
238 xbt_swag_foreach(_var, var_list)
240 lmm_variable_t var = static_cast<lmm_variable_t>(_var);
241 if (var->sharing_weight && var->bound >= 0) {
242 XBT_DEBUG("Working on var (%p)", var);
243 var->new_mu = new_mu(var);
244 XBT_DEBUG("Updating mu : var->mu (%p) : %1.20f -> %1.20f", var, var->mu, var->new_mu);
245 var->mu = var->new_mu;
247 double new_obj = dual_objective(var_list, cnst_list);
248 XBT_DEBUG("Improvement for Objective (%g -> %g) : %g", obj, new_obj, obj - new_obj);
249 xbt_assert(obj - new_obj >= -epsilon_min_error, "Our gradient sucks! (%1.20f)", obj - new_obj);
254 /* Improve the value of lambda_i */
255 xbt_swag_foreach(_cnst, cnst_list)
257 lmm_constraint_t cnst = static_cast<lmm_constraint_t>(_cnst);
258 XBT_DEBUG("Working on cnst (%p)", cnst);
259 cnst->new_lambda = dichotomy(cnst->lambda, partial_diff_lambda, cnst, dichotomy_min_error);
260 XBT_DEBUG("Updating lambda : cnst->lambda (%p) : %1.20f -> %1.20f", cnst, cnst->lambda, cnst->new_lambda);
261 cnst->lambda = cnst->new_lambda;
263 double new_obj = dual_objective(var_list, cnst_list);
264 XBT_DEBUG("Improvement for Objective (%g -> %g) : %g", obj, new_obj, obj - new_obj);
265 xbt_assert(obj - new_obj >= -epsilon_min_error, "Our gradient sucks! (%1.20f)", obj - new_obj);
269 /* Now computes the values of each variable (\rho) based on the values of \lambda and \mu. */
270 XBT_DEBUG("-------------- Check convergence ----------");
271 overall_modification = 0;
272 xbt_swag_foreach(_var, var_list)
274 lmm_variable_t var = static_cast<lmm_variable_t>(_var);
275 if (var->sharing_weight <= 0)
278 double tmp = new_value(var);
280 overall_modification = std::max(overall_modification, fabs(var->value - tmp));
283 XBT_DEBUG("New value of var (%p) = %e, overall_modification = %e", var, var->value, overall_modification);
287 XBT_DEBUG("-------------- Check feasability ----------");
288 if (not __check_feasible(cnst_list, var_list, 0))
289 overall_modification = 1.0;
290 XBT_DEBUG("Iteration %d: overall_modification : %f", iteration, overall_modification);
293 __check_feasible(cnst_list, var_list, 1);
295 if (overall_modification <= epsilon_min_error) {
296 XBT_DEBUG("The method converges in %d iterations.", iteration);
298 if (iteration >= max_iterations) {
299 XBT_DEBUG("Method reach %d iterations, which is the maximum number of iterations allowed.", iteration);
302 if (XBT_LOG_ISENABLED(surf_lagrange, xbt_log_priority_debug)) {
308 * Returns a double value corresponding to the result of a dichotomy process with respect to a given
309 * variable/constraint (\mu in the case of a variable or \lambda in case of a constraint) and a initial value init.
311 * @param init initial value for \mu or \lambda
312 * @param diff a function that computes the differential of with respect a \mu or \lambda
313 * @param var_cnst a pointer to a variable or constraint
314 * @param min_erro a minimum error tolerated
316 * @return a double corresponding to the result of the dichotomy process
318 static double dichotomy(double init, double diff(double, void*), void* var_cnst, double min_error)
322 double overall_error;
329 if (fabs(init) < 1e-20) {
336 diff_0 = diff(1e-16, var_cnst);
338 XBT_CDEBUG(surf_lagrange_dichotomy, "returning 0.0 (diff = %e)", diff_0);
343 double min_diff = diff(min, var_cnst);
344 double max_diff = diff(max, var_cnst);
346 while (overall_error > min_error) {
347 XBT_CDEBUG(surf_lagrange_dichotomy, "[min, max] = [%1.20f, %1.20f] || diffmin, diffmax = %1.20f, %1.20f", min, max,
350 if (min_diff > 0 && max_diff > 0) {
352 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing min");
354 min_diff = diff(min, var_cnst);
356 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing max");
360 } else if (min_diff < 0 && max_diff < 0) {
362 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing max");
364 max_diff = diff(max, var_cnst);
366 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing min");
370 } else if (min_diff < 0 && max_diff > 0) {
371 middle = (max + min) / 2.0;
372 XBT_CDEBUG(surf_lagrange_dichotomy, "Trying (max+min)/2 : %1.20f", middle);
374 if ((fabs(min - middle) < 1e-20) || (fabs(max - middle) < 1e-20)) {
375 XBT_CWARN(surf_lagrange_dichotomy,
376 "Cannot improve the convergence! min=max=middle=%1.20f, diff = %1.20f."
377 " Reaching the 'double' limits. Maybe scaling your function would help ([%1.20f,%1.20f]).",
378 min, max - min, min_diff, max_diff);
381 middle_diff = diff(middle, var_cnst);
383 if (middle_diff < 0) {
384 XBT_CDEBUG(surf_lagrange_dichotomy, "Increasing min");
386 overall_error = max_diff - middle_diff;
387 min_diff = middle_diff;
388 } else if (middle_diff > 0) {
389 XBT_CDEBUG(surf_lagrange_dichotomy, "Decreasing max");
391 overall_error = max_diff - middle_diff;
392 max_diff = middle_diff;
396 } else if (fabs(min_diff) < 1e-20) {
399 } else if (fabs(max_diff) < 1e-20) {
402 } else if (min_diff > 0 && max_diff < 0) {
403 XBT_CWARN(surf_lagrange_dichotomy, "The impossible happened, partial_diff(min) > 0 && partial_diff(max) < 0");
406 XBT_CWARN(surf_lagrange_dichotomy,
407 "diffmin (%1.20f) or diffmax (%1.20f) are something I don't know, taking no action.", min_diff,
413 XBT_CDEBUG(surf_lagrange_dichotomy, "returning %e", (min + max) / 2.0);
415 return ((min + max) / 2.0);
418 static double partial_diff_lambda(double lambda, void* param_cnst)
420 lmm_constraint_t cnst = static_cast<lmm_constraint_t>(param_cnst);
425 XBT_CDEBUG(surf_lagrange_dichotomy, "Computing diff of cnst (%p)", cnst);
427 xbt_swag_t elem_list = &(cnst->enabled_element_set);
429 xbt_swag_foreach(_elem, elem_list)
431 lmm_element_t elem = static_cast<lmm_element_t>(_elem);
432 lmm_variable_t var = elem->variable;
433 xbt_assert(var->sharing_weight > 0);
434 XBT_CDEBUG(surf_lagrange_dichotomy, "Computing sigma_i for var (%p)", var);
435 // Initialize the summation variable
436 double sigma_i = 0.0;
439 for (s_lmm_element_t const& elem : var->cnsts) {
440 sigma_i += elem.constraint->lambda;
443 // add mu_i if this flow has a RTT constraint associated
447 // replace value of cnst->lambda by the value of parameter lambda
448 sigma_i = (sigma_i - cnst->lambda) + lambda;
450 diff += -var->func_fpi(*var, sigma_i);
455 XBT_CDEBUG(surf_lagrange_dichotomy, "d D/d lambda for cnst (%p) at %1.20f = %1.20f", cnst, lambda, diff);
460 /** \brief Attribute the value bound to var->bound.
462 * \param func_fpi inverse of the partial differential of f (f prime inverse, (f')^{-1})
464 * Set default functions to the ones passed as parameters. This is a polymorphism in C pure, enjoy the roots of
468 void lmm_set_default_protocol_function(double (*func_f)(const s_lmm_variable_t& var, double x),
469 double (*func_fp)(const s_lmm_variable_t& var, double x),
470 double (*func_fpi)(const s_lmm_variable_t& var, double x))
473 func_fp_def = func_fp;
474 func_fpi_def = func_fpi;
477 /**************** Vegas and Reno functions *************************/
478 /* NOTE for Reno: all functions consider the network coefficient (alpha) equal to 1. */
481 * For Vegas: $f(x) = \alpha D_f\ln(x)$
482 * Therefore: $fp(x) = \frac{\alpha D_f}{x}$
483 * Therefore: $fpi(x) = \frac{\alpha D_f}{x}$
485 double func_vegas_f(const s_lmm_variable_t& var, double x)
487 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
488 return VEGAS_SCALING * var.sharing_weight * log(x);
491 double func_vegas_fp(const s_lmm_variable_t& var, double x)
493 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
494 return VEGAS_SCALING * var.sharing_weight / x;
497 double func_vegas_fpi(const s_lmm_variable_t& var, double x)
499 xbt_assert(x > 0.0, "Don't call me with stupid values! (%1.20f)", x);
500 return var.sharing_weight / (x / VEGAS_SCALING);
504 * For Reno: $f(x) = \frac{\sqrt{3/2}}{D_f} atan(\sqrt{3/2}D_f x)$
505 * Therefore: $fp(x) = \frac{3}{3 D_f^2 x^2+2}$
506 * Therefore: $fpi(x) = \sqrt{\frac{1}{{D_f}^2 x} - \frac{2}{3{D_f}^2}}$
508 double func_reno_f(const s_lmm_variable_t& var, double x)
510 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
512 return RENO_SCALING * sqrt(3.0 / 2.0) / var.sharing_weight * atan(sqrt(3.0 / 2.0) * var.sharing_weight * x);
515 double func_reno_fp(const s_lmm_variable_t& var, double x)
517 return RENO_SCALING * 3.0 / (3.0 * var.sharing_weight * var.sharing_weight * x * x + 2.0);
520 double func_reno_fpi(const s_lmm_variable_t& var, double x)
524 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
525 xbt_assert(x > 0.0, "Don't call me with stupid values!");
527 res_fpi = 1.0 / (var.sharing_weight * var.sharing_weight * (x / RENO_SCALING)) -
528 2.0 / (3.0 * var.sharing_weight * var.sharing_weight);
531 return sqrt(res_fpi);
534 /* Implementing new Reno-2
535 * For Reno-2: $f(x) = U_f(x_f) = \frac{{2}{D_f}}*ln(2+x*D_f)$
536 * Therefore: $fp(x) = 2/(Weight*x + 2)
537 * Therefore: $fpi(x) = (2*Weight)/x - 4
539 double func_reno2_f(const s_lmm_variable_t& var, double x)
541 xbt_assert(var.sharing_weight > 0.0, "Don't call me with stupid values!");
542 return RENO2_SCALING * (1.0 / var.sharing_weight) *
543 log((x * var.sharing_weight) / (2.0 * x * var.sharing_weight + 3.0));
546 double func_reno2_fp(const s_lmm_variable_t& var, double x)
548 return RENO2_SCALING * 3.0 / (var.sharing_weight * x * (2.0 * var.sharing_weight * x + 3.0));
551 double func_reno2_fpi(const s_lmm_variable_t& var, double x)
553 xbt_assert(x > 0.0, "Don't call me with stupid values!");
554 double tmp = x * var.sharing_weight * var.sharing_weight;
555 double res_fpi = tmp * (9.0 * x + 24.0);
560 res_fpi = RENO2_SCALING * (-3.0 * tmp + sqrt(res_fpi)) / (4.0 * tmp);